08:31 <tomeu> lynxeye: austriancoder : I'm seeing much more flakiness when using mainline, and have found that it is something that appeared between 5.19 and 6.0

08:31 <tomeu> have looked at what changed in etnaviv, drm/sched and meson, and haven't seen anything suspicious

08:32 <tomeu> would you have any ideas?

09:36 <lynxeye> tomeu: How does that flakiness manifest itself? More failures when running things concurrently, or something else?

09:36 <tomeu> yep, that's it

09:37 <tomeu> I'm running with some expected failures that cause the GPU to hang

09:44 <lynxeye> tomeu: Just sent you a patch (mailinglist CC'ed) which might improve things in your case, but be aware that it only swaps one bug with another.

09:47 <tomeu> thanks, will give it a look

10:00 <tomeu> ok, now I know why my 80min bisection was useless: I was carrying a revert of that commit by mistake, so all revisions were good...

10:07 <tomeu> no, it wasn't that

10:07 <tomeu> I so hate bisecting flakiness...

14:01 <tomeu> lynxeye: so indeed, looks like reverting 45ecaea73883 drm/sched: Partial revert of 'drm/sched: Keep s_fence->parent pointer' gets things stable

14:03 <tomeu> not 100%, but 2 flakes out of 1000 is good enough for CI

14:04 <tomeu> guess I could skip the tests that cause GPU hangs and get on with this, if the problems are the timeouts...

14:05 <tomeu> btw, I'm mounting the kernel driver with hw_job_limit=1 job_hang_limit=0

14:05 <tomeu> as I see greater instability with more than one job queued at the same time

14:12 <tomeu> hmm, but there is still something wrong in the power domains that causes that little flakiness

14:12 <tomeu> Pass: 3294, ExpectedFail: 490, Skip: 267, Duration: 3:04, Remaining: 0

14:13 <tomeu> this is if I hack the power up sequence by just directly writing to the registers, instead of letting the power domain code do it

14:17 <tomeu> TBC, I'm referring to the two top commits at https://gitlab.freedesktop.org/tomeu/linux/-/commits/v6.7-for-mesa-ci/?ref_type=heads